Abstract: As an indispensable intelligent technology for advancing Industry 4.0 and the new wave of technological revolution, digital twin technology has garnered significant attention in the field of intelligent mining. Limited by the contradiction between the scale of numerical simulation and computational performance, it is currently difficult to synchronize the structural mechanical response states of fully mechanized mining equipment with their digital twin models in real time. Taking the middle trough of a scraper conveyor as an example, this paper proposes a machine learning-based digital twin modeling method for structural responses. The node responses of the middle trough under different load conditions are obtained through finite element analysis. A hierarchical clustering method is employed to cluster nodes with similar numerical values. A deep neural network (DNN) is then utilized to predict the clustering results and cluster center values of nodes under various load conditions. The predicted cluster center values are used to replace the values of all nodes within the cluster domain. Finally, the global mechanical response state of the middle trough is reconstructed based on the node coordinates and predicted node values. A visualization interface for the digital twin model of the scraper conveyor was developed based on Unity. By deploying sensors to collect load information from the scraper conveyor, the sensor-acquired data is used to drive the DNN in real time, predicting the global deformation and stress responses of the middle trough under different load conditions. This enables the synchronization of the mechanical responses between the digital twin model of the middle trough and its physical counterpart. The research results demonstrate that the time required for the DNN to predict all nodes and complete the 3D node cloud reconstruction is 0.32 seconds, with maximum prediction errors for stress and displacement of 0.97 MPa and 1.98×10⁻³ mm, respectively. The constructed digital twin model is capable of continuously predicting the stress distribution of the middle trough based on signals collected by sensors. The maximum relative error between the predicted stress results at the test points of the middle trough and the experimentally measured values is 33.31%. This verifies the feasibility of the machine learning-based digital twin model for the structural response of the middle trough, providing a new method for the condition monitoring of scraper conveyors.